src/gromacs/ewald/pme_only.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2013,2014,2015,2016,2017 by the GROMACS development team.
   7  * Copyright (c) 2018,2019,2020, by the GROMACS development team, led by
   8  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   9  * and including many others, as listed in the AUTHORS file in the
  10  * top-level source directory and at http://www.gromacs.org.
  11  *
  12  * GROMACS is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU Lesser General Public License
  14  * as published by the Free Software Foundation; either version 2.1
  15  * of the License, or (at your option) any later version.
  16  *
  17  * GROMACS is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  20  * Lesser General Public License for more details.
  21  *
  22  * You should have received a copy of the GNU Lesser General Public
  23  * License along with GROMACS; if not, see
  24  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  25  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  26  *
  27  * If you want to redistribute modifications to GROMACS, please
  28  * consider that scientific software is very special. Version
  29  * control is crucial - bugs must be traceable. We will be happy to
  30  * consider code for inclusion in the official distribution, but
  31  * derived work must not be called official GROMACS. Details are found
  32  * in the README & COPYING files - if they are missing, get the
  33  * official version at http://www.gromacs.org.
  34  *
  35  * To help us fund GROMACS development, we humbly ask that you cite
  36  * the research papers on the package. Check out http://www.gromacs.org.
  37  */
  38 /* IMPORTANT FOR DEVELOPERS:
  39  *
  40  * Triclinic pme stuff isn't entirely trivial, and we've experienced
  41  * some bugs during development (many of them due to me). To avoid
  42  * this in the future, please check the following things if you make
  43  * changes in this file:
  44  *
  45  * 1. You should obtain identical (at least to the PME precision)
  46  *    energies, forces, and virial for
  47  *    a rectangular box and a triclinic one where the z (or y) axis is
  48  *    tilted a whole box side. For instance you could use these boxes:
  49  *
  50  *    rectangular       triclinic
  51  *     2  0  0           2  0  0
  52  *     0  2  0           0  2  0
  53  *     0  0  6           2  2  6
  54  *
  55  * 2. You should check the energy conservation in a triclinic box.
  56  *
  57  * It might seem an overkill, but better safe than sorry.
  58  * /Erik 001109
  59  */
  60
  61 #include "gmxpre.h"
  62
  63 #include "pme_only.h"
  64
  65 #include "config.h"
  66
  67 #include <cassert>
  68 #include <cmath>
  69 #include <cstdio>
  70 #include <cstdlib>
  71 #include <cstring>
  72
  73 #include <memory>
  74 #include <numeric>
  75 #include <vector>
  76
  77 #include "gromacs/domdec/domdec.h"
  78 #include "gromacs/ewald/pme.h"
  79 #include "gromacs/ewald/pme_coordinate_receiver_gpu.h"
  80 #include "gromacs/ewald/pme_force_sender_gpu.h"
  81 #include "gromacs/fft/parallel_3dfft.h"
  82 #include "gromacs/fileio/pdbio.h"
  83 #include "gromacs/gmxlib/network.h"
  84 #include "gromacs/gmxlib/nrnb.h"
  85 #include "gromacs/gpu_utils/hostallocator.h"
  86 #include "gromacs/math/gmxcomplex.h"
  87 #include "gromacs/math/units.h"
  88 #include "gromacs/math/vec.h"
  89 #include "gromacs/mdtypes/commrec.h"
  90 #include "gromacs/mdtypes/forceoutput.h"
  91 #include "gromacs/mdtypes/inputrec.h"
  92 #include "gromacs/mdtypes/simulation_workload.h"
  93 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
  94 #include "gromacs/timing/cyclecounter.h"
  95 #include "gromacs/timing/wallcycle.h"
  96 #include "gromacs/utility/fatalerror.h"
  97 #include "gromacs/utility/futil.h"
  98 #include "gromacs/utility/gmxmpi.h"
  99 #include "gromacs/utility/gmxomp.h"
 100 #include "gromacs/utility/smalloc.h"
 101
 102 #include "pme_gpu_internal.h"
 103 #include "pme_output.h"
 104 #include "pme_pp_communication.h"
 105
 106 /*! \brief environment variable to enable GPU P2P communication */
 107 static const bool c_enableGpuPmePpComms =
 108         (getenv("GMX_GPU_PME_PP_COMMS") != nullptr) && GMX_THREAD_MPI && (GMX_GPU == GMX_GPU_CUDA);
 109
 110 /*! \brief Master PP-PME communication data structure */
 111 struct gmx_pme_pp
 112 {
 113     MPI_Comm             mpi_comm_mysim; /**< MPI communicator for this simulation */
 114     std::vector<PpRanks> ppRanks;        /**< The PP partner ranks                 */
 115     int                  peerRankId;     /**< The peer PP rank id                  */
 116     //@{
 117     /**< Vectors of A- and B-state parameters used to transfer vectors to PME ranks  */
 118     gmx::PaddedHostVector<real> chargeA;
 119     std::vector<real>           chargeB;
 120     std::vector<real>           sqrt_c6A;
 121     std::vector<real>           sqrt_c6B;
 122     std::vector<real>           sigmaA;
 123     std::vector<real>           sigmaB;
 124     //@}
 125     gmx::HostVector<gmx::RVec> x; /**< Vector of atom coordinates to transfer to PME ranks */
 126     std::vector<gmx::RVec>     f; /**< Vector of atom forces received from PME ranks */
 127     //@{
 128     /**< Vectors of MPI objects used in non-blocking communication between multiple PP ranks per PME rank */
 129     std::vector<MPI_Request> req;
 130     std::vector<MPI_Status>  stat;
 131     //@}
 132
 133     /*! \brief object for receiving coordinates using communications operating on GPU memory space */
 134     std::unique_ptr<gmx::PmeCoordinateReceiverGpu> pmeCoordinateReceiverGpu;
 135     /*! \brief object for sending PME force using communications operating on GPU memory space */
 136     std::unique_ptr<gmx::PmeForceSenderGpu> pmeForceSenderGpu;
 137
 138     /*! \brief whether GPU direct communications are active for PME-PP transfers */
 139     bool useGpuDirectComm = false;
 140 };
 141
 142 /*! \brief Initialize the PME-only side of the PME <-> PP communication */
 143 static std::unique_ptr<gmx_pme_pp> gmx_pme_pp_init(const t_commrec* cr)
 144 {
 145     auto pme_pp = std::make_unique<gmx_pme_pp>();
 146
 147 #if GMX_MPI
 148     int rank;
 149
 150     pme_pp->mpi_comm_mysim = cr->mpi_comm_mysim;
 151     MPI_Comm_rank(cr->mpi_comm_mygroup, &rank);
 152     auto ppRanks = get_pme_ddranks(cr, rank);
 153     pme_pp->ppRanks.reserve(ppRanks.size());
 154     for (const auto& ppRankId : ppRanks)
 155     {
 156         pme_pp->ppRanks.push_back({ ppRankId, 0 });
 157     }
 158     // The peer PP rank is the last one.
 159     pme_pp->peerRankId = pme_pp->ppRanks.back().rankId;
 160     pme_pp->req.resize(eCommType_NR * pme_pp->ppRanks.size());
 161     pme_pp->stat.resize(eCommType_NR * pme_pp->ppRanks.size());
 162 #else
 163     GMX_UNUSED_VALUE(cr);
 164 #endif
 165
 166     return pme_pp;
 167 }
 168
 169 static void reset_pmeonly_counters(gmx_wallcycle_t           wcycle,
 170                                    gmx_walltime_accounting_t walltime_accounting,
 171                                    t_nrnb*                   nrnb,
 172                                    int64_t                   step,
 173                                    bool                      useGpuForPme)
 174 {
 175     /* Reset all the counters related to performance over the run */
 176     wallcycle_stop(wcycle, ewcRUN);
 177     wallcycle_reset_all(wcycle);
 178     *nrnb = { 0 };
 179     wallcycle_start(wcycle, ewcRUN);
 180     walltime_accounting_reset_time(walltime_accounting, step);
 181
 182     if (useGpuForPme)
 183     {
 184         resetGpuProfiler();
 185     }
 186 }
 187
 188 static gmx_pme_t* gmx_pmeonly_switch(std::vector<gmx_pme_t*>* pmedata,
 189                                      const ivec               grid_size,
 190                                      real                     ewaldcoeff_q,
 191                                      real                     ewaldcoeff_lj,
 192                                      const t_commrec*         cr,
 193                                      const t_inputrec*        ir)
 194 {
 195     GMX_ASSERT(pmedata, "Bad PME tuning list pointer");
 196     for (auto& pme : *pmedata)
 197     {
 198         GMX_ASSERT(pme, "Bad PME tuning list element pointer");
 199         if (gmx_pme_grid_matches(*pme, grid_size))
 200         {
 201             /* Here we have found an existing PME data structure that suits us.
 202              * However, in the GPU case, we have to reinitialize it - there's only one GPU structure.
 203              * This should not cause actual GPU reallocations, at least (the allocated buffers are never shrunk).
 204              * So, just some grid size updates in the GPU kernel parameters.
 205              * TODO: this should be something like gmx_pme_update_split_params()
 206              */
 207             gmx_pme_reinit(&pme, cr, pme, ir, grid_size, ewaldcoeff_q, ewaldcoeff_lj);
 208             return pme;
 209         }
 210     }
 211
 212     const auto& pme          = pmedata->back();
 213     gmx_pme_t*  newStructure = nullptr;
 214     // Copy last structure with new grid params
 215     gmx_pme_reinit(&newStructure, cr, pme, ir, grid_size, ewaldcoeff_q, ewaldcoeff_lj);
 216     pmedata->push_back(newStructure);
 217     return newStructure;
 218 }
 219
 220 /*! \brief Called by PME-only ranks to receive coefficients and coordinates
 221  *
 222  * \param[in] pme           PME data structure.
 223  * \param[in,out] pme_pp    PME-PP communication structure.
 224  * \param[out] natoms       Number of received atoms.
 225  * \param[out] box        System box, if received.
 226  * \param[out] maxshift_x        Maximum shift in X direction, if received.
 227  * \param[out] maxshift_y        Maximum shift in Y direction, if received.
 228  * \param[out] lambda_q         Free-energy lambda for electrostatics, if received.
 229  * \param[out] lambda_lj         Free-energy lambda for Lennard-Jones, if received.
 230  * \param[out] bEnerVir          Set to true if this is an energy/virial calculation step, otherwise
 231  * set to false. \param[out] step              MD integration step number. \param[out] grid_size PME
 232  * grid size, if received. \param[out] ewaldcoeff_q         Ewald cut-off parameter for
 233  * electrostatics, if received. \param[out] ewaldcoeff_lj         Ewald cut-off parameter for
 234  * Lennard-Jones, if received. \param[in] useGpuForPme      flag on whether PME is on GPU \param[in]
 235  * stateGpu          GPU state propagator object \param[in] runMode           PME run mode
 236  *
 237  * \retval pmerecvqxX             All parameters were set, chargeA and chargeB can be NULL.
 238  * \retval pmerecvqxFINISH        No parameters were set.
 239  * \retval pmerecvqxSWITCHGRID    Only grid_size and *ewaldcoeff were set.
 240  * \retval pmerecvqxRESETCOUNTERS *step was set.
 241  */
 242 static int gmx_pme_recv_coeffs_coords(struct gmx_pme_t*            pme,
 243                                       gmx_pme_pp*                  pme_pp,
 244                                       int*                         natoms,
 245                                       matrix                       box,
 246                                       int*                         maxshift_x,
 247                                       int*                         maxshift_y,
 248                                       real*                        lambda_q,
 249                                       real*                        lambda_lj,
 250                                       gmx_bool*                    bEnerVir,
 251                                       int64_t*                     step,
 252                                       ivec*                        grid_size,
 253                                       real*                        ewaldcoeff_q,
 254                                       real*                        ewaldcoeff_lj,
 255                                       bool                         useGpuForPme,
 256                                       gmx::StatePropagatorDataGpu* stateGpu,
 257                                       PmeRunMode gmx_unused runMode)
 258 {
 259     int status = -1;
 260     int nat    = 0;
 261
 262 #if GMX_MPI
 263     unsigned int flags          = 0;
 264     int          messages       = 0;
 265     bool         atomSetChanged = false;
 266
 267     do
 268     {
 269         gmx_pme_comm_n_box_t cnb;
 270         cnb.flags = 0;
 271
 272         /* Receive the send count, box and time step from the peer PP node */
 273         MPI_Recv(&cnb, sizeof(cnb), MPI_BYTE, pme_pp->peerRankId, eCommType_CNB,
 274                  pme_pp->mpi_comm_mysim, MPI_STATUS_IGNORE);
 275
 276         /* We accumulate all received flags */
 277         flags |= cnb.flags;
 278
 279         *step = cnb.step;
 280
 281         if (debug)
 282         {
 283             fprintf(debug, "PME only rank receiving:%s%s%s%s%s\n",
 284                     (cnb.flags & PP_PME_CHARGE) ? " charges" : "",
 285                     (cnb.flags & PP_PME_COORD) ? " coordinates" : "",
 286                     (cnb.flags & PP_PME_FINISH) ? " finish" : "",
 287                     (cnb.flags & PP_PME_SWITCHGRID) ? " switch grid" : "",
 288                     (cnb.flags & PP_PME_RESETCOUNTERS) ? " reset counters" : "");
 289         }
 290
 291         pme_pp->useGpuDirectComm = ((cnb.flags & PP_PME_GPUCOMMS) != 0);
 292         GMX_ASSERT(!pme_pp->useGpuDirectComm || (pme_pp->pmeForceSenderGpu != nullptr),
 293                    "The use of GPU direct communication for PME-PP is enabled, "
 294                    "but the PME GPU force reciever object does not exist");
 295
 296         if (cnb.flags & PP_PME_FINISH)
 297         {
 298             status = pmerecvqxFINISH;
 299         }
 300
 301         if (cnb.flags & PP_PME_SWITCHGRID)
 302         {
 303             /* Special case, receive the new parameters and return */
 304             copy_ivec(cnb.grid_size, *grid_size);
 305             *ewaldcoeff_q  = cnb.ewaldcoeff_q;
 306             *ewaldcoeff_lj = cnb.ewaldcoeff_lj;
 307
 308             status = pmerecvqxSWITCHGRID;
 309         }
 310
 311         if (cnb.flags & PP_PME_RESETCOUNTERS)
 312         {
 313             /* Special case, receive the step (set above) and return */
 314             status = pmerecvqxRESETCOUNTERS;
 315         }
 316
 317         if (cnb.flags & (PP_PME_CHARGE | PP_PME_SQRTC6 | PP_PME_SIGMA))
 318         {
 319             atomSetChanged = true;
 320
 321             /* Receive the send counts from the other PP nodes */
 322             for (auto& sender : pme_pp->ppRanks)
 323             {
 324                 if (sender.rankId == pme_pp->peerRankId)
 325                 {
 326                     sender.numAtoms = cnb.natoms;
 327                 }
 328                 else
 329                 {
 330                     MPI_Irecv(&sender.numAtoms, sizeof(sender.numAtoms), MPI_BYTE, sender.rankId,
 331                               eCommType_CNB, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 332                 }
 333             }
 334             MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 335             messages = 0;
 336
 337             nat = 0;
 338             for (const auto& sender : pme_pp->ppRanks)
 339             {
 340                 nat += sender.numAtoms;
 341             }
 342
 343             if (cnb.flags & PP_PME_CHARGE)
 344             {
 345                 pme_pp->chargeA.resizeWithPadding(nat);
 346             }
 347             if (cnb.flags & PP_PME_CHARGEB)
 348             {
 349                 pme_pp->chargeB.resize(nat);
 350             }
 351             if (cnb.flags & PP_PME_SQRTC6)
 352             {
 353                 pme_pp->sqrt_c6A.resize(nat);
 354             }
 355             if (cnb.flags & PP_PME_SQRTC6B)
 356             {
 357                 pme_pp->sqrt_c6B.resize(nat);
 358             }
 359             if (cnb.flags & PP_PME_SIGMA)
 360             {
 361                 pme_pp->sigmaA.resize(nat);
 362             }
 363             if (cnb.flags & PP_PME_SIGMAB)
 364             {
 365                 pme_pp->sigmaB.resize(nat);
 366             }
 367             pme_pp->x.resize(nat);
 368             pme_pp->f.resize(nat);
 369
 370             /* maxshift is sent when the charges are sent */
 371             *maxshift_x = cnb.maxshift_x;
 372             *maxshift_y = cnb.maxshift_y;
 373
 374             /* Receive the charges in place */
 375             for (int q = 0; q < eCommType_NR; q++)
 376             {
 377                 real* bufferPtr;
 378
 379                 if (!(cnb.flags & (PP_PME_CHARGE << q)))
 380                 {
 381                     continue;
 382                 }
 383                 switch (q)
 384                 {
 385                     case eCommType_ChargeA: bufferPtr = pme_pp->chargeA.data(); break;
 386                     case eCommType_ChargeB: bufferPtr = pme_pp->chargeB.data(); break;
 387                     case eCommType_SQRTC6A: bufferPtr = pme_pp->sqrt_c6A.data(); break;
 388                     case eCommType_SQRTC6B: bufferPtr = pme_pp->sqrt_c6B.data(); break;
 389                     case eCommType_SigmaA: bufferPtr = pme_pp->sigmaA.data(); break;
 390                     case eCommType_SigmaB: bufferPtr = pme_pp->sigmaB.data(); break;
 391                     default: gmx_incons("Wrong eCommType");
 392                 }
 393                 nat = 0;
 394                 for (const auto& sender : pme_pp->ppRanks)
 395                 {
 396                     if (sender.numAtoms > 0)
 397                     {
 398                         MPI_Irecv(bufferPtr + nat, sender.numAtoms * sizeof(real), MPI_BYTE,
 399                                   sender.rankId, q, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 400                         nat += sender.numAtoms;
 401                         if (debug)
 402                         {
 403                             fprintf(debug, "Received from PP rank %d: %d %s\n", sender.rankId,
 404                                     sender.numAtoms,
 405                                     (q == eCommType_ChargeA || q == eCommType_ChargeB) ? "charges"
 406                                                                                        : "params");
 407                         }
 408                     }
 409                 }
 410             }
 411         }
 412
 413         if (cnb.flags & PP_PME_COORD)
 414         {
 415             if (atomSetChanged)
 416             {
 417                 gmx_pme_reinit_atoms(pme, nat, pme_pp->chargeA.data());
 418                 if (useGpuForPme)
 419                 {
 420                     stateGpu->reinit(nat, nat);
 421                     pme_gpu_set_device_x(pme, stateGpu->getCoordinates());
 422                 }
 423                 if (pme_pp->useGpuDirectComm)
 424                 {
 425                     GMX_ASSERT(runMode == PmeRunMode::GPU,
 426                                "GPU Direct PME-PP communication has been enabled, "
 427                                "but PME run mode is not PmeRunMode::GPU\n");
 428
 429                     // This rank will have its data accessed directly by PP rank, so needs to send the remote addresses.
 430                     pme_pp->pmeCoordinateReceiverGpu->sendCoordinateBufferAddressToPpRanks(
 431                             stateGpu->getCoordinates());
 432                     pme_pp->pmeForceSenderGpu->sendForceBufferAddressToPpRanks(
 433                             reinterpret_cast<rvec*>(pme_gpu_get_device_f(pme)));
 434                 }
 435             }
 436
 437
 438             /* The box, FE flag and lambda are sent along with the coordinates
 439              *  */
 440             copy_mat(cnb.box, box);
 441             *lambda_q  = cnb.lambda_q;
 442             *lambda_lj = cnb.lambda_lj;
 443             *bEnerVir  = ((cnb.flags & PP_PME_ENER_VIR) != 0U);
 444             *step      = cnb.step;
 445
 446             /* Receive the coordinates in place */
 447             nat = 0;
 448             for (const auto& sender : pme_pp->ppRanks)
 449             {
 450                 if (sender.numAtoms > 0)
 451                 {
 452                     if (pme_pp->useGpuDirectComm)
 453                     {
 454                         pme_pp->pmeCoordinateReceiverGpu->launchReceiveCoordinatesFromPpCudaDirect(
 455                                 sender.rankId);
 456                     }
 457                     else
 458                     {
 459                         MPI_Irecv(pme_pp->x[nat], sender.numAtoms * sizeof(rvec), MPI_BYTE, sender.rankId,
 460                                   eCommType_COORD, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
 461                     }
 462                     nat += sender.numAtoms;
 463                     if (debug)
 464                     {
 465                         fprintf(debug,
 466                                 "Received from PP rank %d: %d "
 467                                 "coordinates\n",
 468                                 sender.rankId, sender.numAtoms);
 469                     }
 470                 }
 471             }
 472
 473             if (pme_pp->useGpuDirectComm)
 474             {
 475                 pme_pp->pmeCoordinateReceiverGpu->enqueueWaitReceiveCoordinatesFromPpCudaDirect();
 476             }
 477
 478             status = pmerecvqxX;
 479         }
 480
 481         /* Wait for the coordinates and/or charges to arrive */
 482         MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 483         messages = 0;
 484     } while (status == -1);
 485 #else
 486     GMX_UNUSED_VALUE(pme);
 487     GMX_UNUSED_VALUE(pme_pp);
 488     GMX_UNUSED_VALUE(box);
 489     GMX_UNUSED_VALUE(maxshift_x);
 490     GMX_UNUSED_VALUE(maxshift_y);
 491     GMX_UNUSED_VALUE(lambda_q);
 492     GMX_UNUSED_VALUE(lambda_lj);
 493     GMX_UNUSED_VALUE(bEnerVir);
 494     GMX_UNUSED_VALUE(step);
 495     GMX_UNUSED_VALUE(grid_size);
 496     GMX_UNUSED_VALUE(ewaldcoeff_q);
 497     GMX_UNUSED_VALUE(ewaldcoeff_lj);
 498     GMX_UNUSED_VALUE(useGpuForPme);
 499     GMX_UNUSED_VALUE(stateGpu);
 500
 501     status = pmerecvqxX;
 502 #endif
 503
 504     if (status == pmerecvqxX)
 505     {
 506         *natoms = nat;
 507     }
 508
 509     return status;
 510 }
 511
 512 #if GMX_MPI
 513 /*! \brief Send force data to PP ranks */
 514 static void sendFToPP(void* sendbuf, PpRanks receiver, gmx_pme_pp* pme_pp, int* messages)
 515 {
 516
 517     if (pme_pp->useGpuDirectComm)
 518     {
 519         GMX_ASSERT((pme_pp->pmeForceSenderGpu != nullptr),
 520                    "The use of GPU direct communication for PME-PP is enabled, "
 521                    "but the PME GPU force reciever object does not exist");
 522
 523         pme_pp->pmeForceSenderGpu->sendFToPpCudaDirect(receiver.rankId);
 524     }
 525     else
 526     {
 527         // Send using MPI
 528         MPI_Isend(sendbuf, receiver.numAtoms * sizeof(rvec), MPI_BYTE, receiver.rankId, 0,
 529                   pme_pp->mpi_comm_mysim, &pme_pp->req[*messages]);
 530         *messages = *messages + 1;
 531     }
 532 }
 533 #endif
 534
 535 /*! \brief Send the PME mesh force, virial and energy to the PP-only ranks. */
 536 static void gmx_pme_send_force_vir_ener(const gmx_pme_t& pme,
 537                                         gmx_pme_pp*      pme_pp,
 538                                         const PmeOutput& output,
 539                                         real             dvdlambda_q,
 540                                         real             dvdlambda_lj,
 541                                         float            cycles)
 542 {
 543 #if GMX_MPI
 544     gmx_pme_comm_vir_ene_t cve;
 545     int                    messages, ind_start, ind_end;
 546     cve.cycles = cycles;
 547
 548     /* Now the evaluated forces have to be transferred to the PP nodes */
 549     messages = 0;
 550     ind_end  = 0;
 551     for (const auto& receiver : pme_pp->ppRanks)
 552     {
 553         ind_start     = ind_end;
 554         ind_end       = ind_start + receiver.numAtoms;
 555         void* sendbuf = const_cast<void*>(static_cast<const void*>(output.forces_[ind_start]));
 556         if (pme_pp->useGpuDirectComm)
 557         {
 558             // Data will be transferred directly from GPU.
 559             rvec* d_f = reinterpret_cast<rvec*>(pme_gpu_get_device_f(&pme));
 560             sendbuf   = reinterpret_cast<void*>(&d_f[ind_start]);
 561         }
 562         sendFToPP(sendbuf, receiver, pme_pp, &messages);
 563     }
 564
 565     /* send virial and energy to our last PP node */
 566     copy_mat(output.coulombVirial_, cve.vir_q);
 567     copy_mat(output.lennardJonesVirial_, cve.vir_lj);
 568     cve.energy_q     = output.coulombEnergy_;
 569     cve.energy_lj    = output.lennardJonesEnergy_;
 570     cve.dvdlambda_q  = dvdlambda_q;
 571     cve.dvdlambda_lj = dvdlambda_lj;
 572     /* check for the signals to send back to a PP node */
 573     cve.stop_cond = gmx_get_stop_condition();
 574
 575     cve.cycles = cycles;
 576
 577     if (debug)
 578     {
 579         fprintf(debug, "PME rank sending to PP rank %d: virial and energy\n", pme_pp->peerRankId);
 580     }
 581     MPI_Isend(&cve, sizeof(cve), MPI_BYTE, pme_pp->peerRankId, 1, pme_pp->mpi_comm_mysim,
 582               &pme_pp->req[messages++]);
 583
 584     /* Wait for the forces to arrive */
 585     MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
 586 #else
 587     gmx_call("MPI not enabled");
 588     GMX_UNUSED_VALUE(pme);
 589     GMX_UNUSED_VALUE(pme_pp);
 590     GMX_UNUSED_VALUE(output);
 591     GMX_UNUSED_VALUE(dvdlambda_q);
 592     GMX_UNUSED_VALUE(dvdlambda_lj);
 593     GMX_UNUSED_VALUE(cycles);
 594 #endif
 595 }
 596
 597 int gmx_pmeonly(struct gmx_pme_t*         pme,
 598                 const t_commrec*          cr,
 599                 t_nrnb*                   mynrnb,
 600                 gmx_wallcycle*            wcycle,
 601                 gmx_walltime_accounting_t walltime_accounting,
 602                 t_inputrec*               ir,
 603                 PmeRunMode                runMode)
 604 {
 605     int      ret;
 606     int      natoms = 0;
 607     matrix   box;
 608     real     lambda_q   = 0;
 609     real     lambda_lj  = 0;
 610     int      maxshift_x = 0, maxshift_y = 0;
 611     real     dvdlambda_q, dvdlambda_lj;
 612     float    cycles;
 613     int      count;
 614     gmx_bool bEnerVir = FALSE;
 615     int64_t  step;
 616
 617     /* This data will only use with PME tuning, i.e. switching PME grids */
 618     std::vector<gmx_pme_t*> pmedata;
 619     pmedata.push_back(pme);
 620
 621     auto pme_pp = gmx_pme_pp_init(cr);
 622
 623     std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
 624     // TODO the variable below should be queried from the task assignment info
 625     const bool useGpuForPme = (runMode == PmeRunMode::GPU) || (runMode == PmeRunMode::Mixed);
 626     if (useGpuForPme)
 627     {
 628         const void* commandStream = pme_gpu_get_device_stream(pme);
 629         const void* deviceContext = pme_gpu_get_device_context(pme);
 630
 631         changePinningPolicy(&pme_pp->chargeA, pme_get_pinning_policy());
 632         changePinningPolicy(&pme_pp->x, pme_get_pinning_policy());
 633         if (c_enableGpuPmePpComms)
 634         {
 635             pme_pp->pmeCoordinateReceiverGpu = std::make_unique<gmx::PmeCoordinateReceiverGpu>(
 636                     commandStream, pme_pp->mpi_comm_mysim, pme_pp->ppRanks);
 637             pme_pp->pmeForceSenderGpu = std::make_unique<gmx::PmeForceSenderGpu>(
 638                     commandStream, pme_pp->mpi_comm_mysim, pme_pp->ppRanks);
 639         }
 640         // TODO: Special PME-only constructor is used here. There is no mechanism to prevent from using the other constructor here.
 641         //       This should be made safer.
 642         stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
 643                 commandStream, deviceContext, GpuApiCallBehavior::Async,
 644                 pme_gpu_get_padding_size(pme), wcycle);
 645     }
 646
 647     clear_nrnb(mynrnb);
 648
 649     count = 0;
 650     do /****** this is a quasi-loop over time steps! */
 651     {
 652         /* The reason for having a loop here is PME grid tuning/switching */
 653         do
 654         {
 655             /* Domain decomposition */
 656             ivec newGridSize;
 657             real ewaldcoeff_q = 0, ewaldcoeff_lj = 0;
 658             ret = gmx_pme_recv_coeffs_coords(pme, pme_pp.get(), &natoms, box, &maxshift_x,
 659                                              &maxshift_y, &lambda_q, &lambda_lj, &bEnerVir, &step,
 660                                              &newGridSize, &ewaldcoeff_q, &ewaldcoeff_lj,
 661                                              useGpuForPme, stateGpu.get(), runMode);
 662
 663             if (ret == pmerecvqxSWITCHGRID)
 664             {
 665                 /* Switch the PME grid to newGridSize */
 666                 pme = gmx_pmeonly_switch(&pmedata, newGridSize, ewaldcoeff_q, ewaldcoeff_lj, cr, ir);
 667             }
 668
 669             if (ret == pmerecvqxRESETCOUNTERS)
 670             {
 671                 /* Reset the cycle and flop counters */
 672                 reset_pmeonly_counters(wcycle, walltime_accounting, mynrnb, step, useGpuForPme);
 673             }
 674         } while (ret == pmerecvqxSWITCHGRID || ret == pmerecvqxRESETCOUNTERS);
 675
 676         if (ret == pmerecvqxFINISH)
 677         {
 678             /* We should stop: break out of the loop */
 679             break;
 680         }
 681
 682         if (count == 0)
 683         {
 684             wallcycle_start(wcycle, ewcRUN);
 685             walltime_accounting_start_time(walltime_accounting);
 686         }
 687
 688         wallcycle_start(wcycle, ewcPMEMESH);
 689
 690         dvdlambda_q  = 0;
 691         dvdlambda_lj = 0;
 692
 693         // TODO Make a struct of array refs onto these per-atom fields
 694         // of pme_pp (maybe box, energy and virial, too; and likewise
 695         // from mdatoms for the other call to gmx_pme_do), so we have
 696         // fewer lines of code and less parameter passing.
 697         gmx::StepWorkload stepWork;
 698         stepWork.computeVirial = bEnerVir;
 699         stepWork.computeEnergy = bEnerVir;
 700         stepWork.computeForces = true;
 701         PmeOutput output;
 702         if (useGpuForPme)
 703         {
 704             stepWork.haveDynamicBox      = false;
 705             stepWork.useGpuPmeFReduction = pme_pp->useGpuDirectComm;
 706             // TODO this should be set properly by gmx_pme_recv_coeffs_coords,
 707             // or maybe use inputrecDynamicBox(ir), at the very least - change this when this codepath is tested!
 708             pme_gpu_prepare_computation(pme, box, wcycle, stepWork);
 709             if (!pme_pp->useGpuDirectComm)
 710             {
 711                 stateGpu->copyCoordinatesToGpu(gmx::ArrayRef<gmx::RVec>(pme_pp->x), gmx::AtomLocality::All);
 712             }
 713             // On the separate PME rank we do not need a synchronizer as we schedule everything in a single stream
 714             // TODO: with pme on GPU the receive should make a list of synchronizers and pass it here #3157
 715             auto xReadyOnDevice = nullptr;
 716
 717             pme_gpu_launch_spread(pme, xReadyOnDevice, wcycle);
 718             pme_gpu_launch_complex_transforms(pme, wcycle, stepWork);
 719             pme_gpu_launch_gather(pme, wcycle);
 720             output = pme_gpu_wait_finish_task(pme, bEnerVir, wcycle);
 721             pme_gpu_reinit_computation(pme, wcycle);
 722         }
 723         else
 724         {
 725             GMX_ASSERT(pme_pp->x.size() == static_cast<size_t>(natoms),
 726                        "The coordinate buffer should have size natoms");
 727
 728             gmx_pme_do(pme, pme_pp->x, pme_pp->f, pme_pp->chargeA.data(), pme_pp->chargeB.data(),
 729                        pme_pp->sqrt_c6A.data(), pme_pp->sqrt_c6B.data(), pme_pp->sigmaA.data(),
 730                        pme_pp->sigmaB.data(), box, cr, maxshift_x, maxshift_y, mynrnb, wcycle,
 731                        output.coulombVirial_, output.lennardJonesVirial_, &output.coulombEnergy_,
 732                        &output.lennardJonesEnergy_, lambda_q, lambda_lj, &dvdlambda_q,
 733                        &dvdlambda_lj, stepWork);
 734             output.forces_ = pme_pp->f;
 735         }
 736
 737         cycles = wallcycle_stop(wcycle, ewcPMEMESH);
 738         gmx_pme_send_force_vir_ener(*pme, pme_pp.get(), output, dvdlambda_q, dvdlambda_lj, cycles);
 739
 740         count++;
 741     } /***** end of quasi-loop, we stop with the break above */
 742     while (TRUE);
 743
 744     walltime_accounting_end_time(walltime_accounting);
 745
 746     return 0;
 747 }